Towards Cross-Tokenizer Distillation: the Universal Logit Distillation Loss for LLMs
Paper
•
2402.12030
•
Published
•
3
ULD Loss (Universal LLMs Distillation)
The ULD loss, based on optimal transport, enables distillation across different LLM families without requiring shared tokenizers.
-
-
mistralai/Mistral-7B-Instruct-v0.2
Text Generation • 7B • Updated • 2.78M • • 3k -
meta-llama/Llama-2-7b-chat-hf
Text Generation • 7B • Updated • 421k • 4.63k -
EleutherAI/pythia-160m-deduped
Text Generation • 0.2B • Updated • 36k • 3 -
EleutherAI/pythia-410m-deduped
Text Generation • 0.5B • Updated • 13k • 20 -
EleutherAI/pythia-1b-deduped
Text Generation • 1B • Updated • 6.19k • 19 -
bigscience/bloomz-560m
Text Generation • 0.6B • Updated • 3.19M • 129 -
bigscience/mt0-base
Text Generation • 0.6B • Updated • 699 • 32 -
facebook/opt-350m
Text Generation • Updated • 118k • 148 -
rajpurkar/squad
Viewer • Updated • 98.2k • 85.9k • 340 -
google-research-datasets/qed
Updated • 213 • 3 -
GEM/FairytaleQA
Viewer • Updated • 10.6k • 198 • 9 -
qiaojin/PubMedQA
Viewer • Updated • 274k • 19.3k • 267 -
knkarthick/dialogsum
Viewer • Updated • 14.5k • 12.3k • 222